AI Companies Really, Really Want to Buy Your Work Files and Emails

WIRED | Steven Levy

Artificial intelligence developers including OpenAI and Anthropic are spending tens of billions of dollars purchasing private corporate work files, emails, and internal communication archives to train next-generation models. Data brokers like Mercor, Micro1, and Handshake are driving this acquisition boom, offering payments from $350K to $1.8M to monetize proprietary organizational workflows.

Private brokers claim automated scrubbing eliminates personally identifiable details, yet legal experts and labor organizations warn that unstructured workplace files inevitably leak sensitive employee data. Data brokers now drive startup creation. During Spirit Airlines' bankruptcy liquidation, Micro1 submitted a $12.5 million bid for 80,000 email accounts, 500 million internal messages, and 20.6 million SharePoint files, topping Google's $10 million offer before labor objections stalled the sale. While brokers project data acquisition spending to escalate from $40 billion to $50 billion annually toward $1 trillion, workers face systemic risks as their daily operational outputs are harvested to build automated replacements.

Comment

The commercial harvesting of un-redacted enterprise telemetry reveals the technical boundaries of synthetic data generation for domain-specific Large Language Models. Major developers attempting to train reasoning engines for complex administrative workflows find synthetic training sets incapable of replicating real-world human error, institutional ambiguity, and rapid context switching. The bidding war over Spirit Airlines' 20.6 million Microsoft SharePoint documents demonstrates how raw operational noise has become a primary valuation driver for frontier AI models.

As commercial brokers systematically ingest enterprise communication networks, defensive counter-intelligence protocols across dual-use technology sectors face structural exposure. Adversarial actors exploiting model inversion techniques can extract proprietary business logic and supply chain dependencies embedded within fine-tuned model weights. Consequently, enterprise telemetry harvested from Spirit Airlines' internal servers risks establishing permanent cyber vulnerabilities within downstream commercial and dual-use foundation models.

Strategic Question for Discussion

If enterprise telemetry like Spirit Airlines' Microsoft SharePoint corpus becomes standard training data for dual-use foundation models, which factor poses the greater long-term risk — structural data leakage through model inversion, or the institutional erosion of workplace privacy rights?

The available evidence points toward model inversion and unintended inference as the more severe strategic vulnerability. While privacy violations provoke immediate regulatory pushback, embedded operational telemetry within neural weights creates permanent, un-scrubbed vectors for competitive espionage and industrial reconnaissance. My assessment is that current data de-identification techniques will consistently fall short against sophisticated extraction attacks targeted at specialized enterprise models.

Share your assessment in the comments below.